TenStrip XLPony

XNP:I'm deliberately mis-tagging it as Pony base because functionality and output-wise its a 3D RealPony model that is heavily NAI/Illustrious rooted. Plus, the NAI section is highly degen, and that's coming from me. The base model is capable of so much more than gay anthro cartoons. Does everything 10Love and previous versions can't do, but is worse at realistic humans without lora stacking and has very little (some) HyperLora compatibility. I've seen success at passing base images into 10Love2 as the highres pass model at 8 steps-0.2/0.3 denoise to make humans much more realistic while keeping the base composition and that pass can use HyperLora to change the face. The most consistent prompt setup seems to be score_9, score_8_up, rest of prompt with important tags first, masterpiece. It's noob, so any unhinged random smut idea is probably within reach. Just needs the right prompt and when you're tinkering with prompting you'll notice the model can guide and respond very strongly and there are definitely right prompts and wrong prompts. Noob and Pony see some tags differently, and while the mix is fairly well-married, there can still be unintended association. Prompts like hot; hot is not a beauty token, it will make her sweaty and wet. Score tags definitely seem to clean up a lot without removing elements from the prompt or harming the image. End-load all the random quality tags, and most of them really don't do anything unless they're chained together, outside of masterpiece and highres which affect a lot.LCM 10-14 steps seems to be the best zone. This one is DMD-lite and can easily go to 1.5-1.8 CFG to use negatives but that's pretty untested how functional that is. Extremely good at latent highres passes up to almost 2k at 0.5 denoise for consistency or 0.7 for a more complex remix. There will sometimes be some slight border to the images on one side, idk if that's a noob thing or some kind of bucket error from the merge. Easy to just crop. Some prompts will keep hands and anatomy better than others. You'll almost always want to stay in the normal SDXL buckets for the lower resolution base image. It's not really a furry model noob is just full of anthro stuff, but in this mix furry images will still usually always favor human hands. My 10noob mix seems slightly more adept at full anthro.This noob character database is pretty useful and the base character name tag on it's own can help diversify faces.Pony and base SDXL Loras are only semi-effective and transfer the basic concept. Which can be good, or bad, if you want to hijack the models styling or not. llustrious and NAI loras are too effective and usually hijack the styling completely reverting it to anime. I had many versions with some being a lot more photoreal, but they always lose the noob training when pushed into photo models too much and start to favor extremely blurry details unprompted. Ultimately this was the compromise mix to preserve what Noob can do while making it function more like the models in this family. There is a small lora addition somewhere along the recipe to correct futa and male anatomy which noob+IL merges struggle with at low CFG. The model will sometimes favor male or hermaphroditic anatomy and bulges without any contrary prompting to refer to pussy, 1girl, (female), etc. I'm still working on my ultra-curated photo dataset for my own full finetune. It's only at ~4.7k highres images of only sexy females, POV sex shots, and very feminine trans models with nice poses and aesthetics. I think noob+Pony like this is probably a leading base candidate because there is a lot of nonsense it can slowly unlearn while still keeping many unique outfit and sex prompts. 10Love versions:Full comfy workflow with HyperLora:https://huggingface.co/TenStrip/Workflows/tree/main"Love" from the bigLove models, although this is definitely not a simple A + B merge with them, just meant to operate like them and the other lust models but with a different look. Meth merge of XL models, literally everything under the sink. XLP, lust, Illustrious, Pony, 3dcdg_pony, cunetta, etc. Clean exponential curves were used during merges, no random 0.5 destructive mid block chops. No dirty Lora merges. All that mattered was average output look (my taste and in-line with the other XLP versions), no overly freckled skin, 50%+ success rate for output composition, and works with the Il penis size slider, and futa/otoko images not turning testicles into vaginas. This is a personal use merge by an experienced user, if you don't inpaint or upscale you won't get the most out of them. Don't bother using on site. Very fast, seconds on my card. Made to feed into Wan video img2vid or generate fast 1girl ideas.Solo figure masters. +femboys and tgirls. Probably males if that's your thing. DMD included, use LCM.LCM can suck and usually is very homogenous, but there's 5 previous versions up there that try to bypass LCM to have eulerA turbo'd and ultimately those all didn't work because the usability is too niche. More steps = refinement and detail. 4-8 steps for the lowres amateur/real look. Past 10-12 steps you get into a refined illustrious/real pony look. Save as webp at 50% quality for the compressed image look. Use HyperLora to wildcard face diversity. Pony, XL, loras, with some IL loras working*.Upscale 1.3-1.4x. 3 steps, 0.3 denoise with a good upscaler. Or 6-12 steps, 0.59-0.7 for hires latent overwrite pass.10L usage:2 common issues with these model types:Gratuitous bare breasts and lack of face diversity.For breastsDON'T use large breasts or mention nipples if you don't want them pulled out. Large bust, large cleavage, corset pushup, micropressure, bra, big chest, +breast size sliders or controls are all ways to get clothed big bust.For facesHyperLora. Use faces from flux, loras, video games, Instagram, other outputs you have, w/e. Use at 0.1-0.3 low strength for wildcard or 0.8-1.4 for Lora-like effect. Models were merged to be compatible with the Xl fidelity lora. The lora does impact composition from pulling light data off of the input so use that to controlnet lighting and eyes, or set the workflow to mask and inpaint the face if the lora manipulates the output away from prompt too much. Requires comfy, get the workflow off of the obvious images that use it and it's bareboned and simple. You can change the hyperlora strength by +/- 0.01 to make minute seed adjustments on locked seeds as well. I don't think there's a faster/better nsfw workflow with that much control. Every image of a face is a lora. Obviously you can get into some big trouble with this so keep it tasteful.XLove_V2:Back to Pony. score_9, score_8_up seems very effective for weeding out nonsense. Much more user friendly for prompt exploration. CFG 1.5 if you need negatives. Composition seems similar to lustify sometimes. Adding IL loras and Pony loras goes into uncharted territory so the merge stands on it's own, worth trying out. Loras will start distorting faces and hands so it'll need edits.XLove_V1:Unstable version. Score tags don't seem to help. Masterpiece doesn't either. Illustrious version, and many loras/sliders seem usable, but the downside is having Il with no negatives. Not user friendly. Best bet is to take lustify/biglove prompts and remix them. Being creative leads to monstrosity. POV angle only for porn. Both lustXL/love/asp models and Illustrious have overlapped tags that think differently and the tokens fight each other over 1boy/1girl with angle concepts. Not even sure how to properly prompt for it, but sometimes things work and they work very well outside of 2girls = 3girls, 1girl 1boy = 1girl getting dp. Hands and limbs problematic as well. This one is essentially unusable for multiple figures outside of precise prompts. For user friendly version skip to V2.X1-X4:Baked slightly turbo'd models that are very Pony based. Consult user images/my images. All experiments with turbo loras that aren't dmd2 - lcm and probably not optimal merges because I was learning.X5:Just a worse non-DMD version of the Loves.If you think your model was used in the merge it probably was. 20-30 models used and difference trained out in this lineage so far so shout yourself out. I only save good block weight curves.